Anomaly detection on single variate time series with concept drifts

Abstract

This study introduces an anomaly detection method on time series data. An ensemble LSTM CNN network with normalization and regularization of time series data incorporated with a concept drift adaptation technique is used for profiling the normal of a time series. Dynamic Time Warping(DTW) algorithm is used to generate warp distance between observed time window and profiler output to be used as a distance measure. Anomalies are detected when the distance between observed time window and profiler output exceeds a pre-defined threshold value. This study compares performance of few profiling methods on Numenta Anomaly Detection benchmark data set. Anomaly detection method implemented using ensemble LSTM CNN network with DTW introduced in this study performs better that baseline anomaly detection method implemented using ARMA with DTW. Keywords: anomaly detection, time series, LSTM, CNN, ensemble networks, concept drift, dynamic time warping